search

cambridge-mlg / cnaps

Code for: "Fast and Flexible Multi-Task Classification Using Conditional Neural Adaptive Processes" and "TaskNorm: Rethinking Batch Normalization for Meta-Learning"
MIT License
159 stars 22 forks source link

Meta-Dataset version #12

Closed xialeiliu closed 3 years ago

xialeiliu commented 3 years ago

Could you clarify which meta-dataset version is used for this repo? Since Meta-dataset repo is updated very frequently.

jfb54 commented 3 years ago

We try to keep up with the latest Meta-dataset version. We updated CNAPs to work with Meta-dataset version on 24 December 2020. Sorry, I don't have the precise Meta-dataset version number. Let us know if something doesn't work with the latest Meta-dataset and we'll fix anything promptly. We only use a small part of their code (the dataset reader). Next time we do an update, we will publish the precise version of Meta-dataset that we last tested with.

xialeiliu commented 3 years ago

I see. Thanks. I just updated a couple of lines in .gin file in order to be compatible with the latest meta-dataset version. It worked.

jfb54 commented 3 years ago

Thanks. Just curious, what lines did you need to change? We can then make the fix in the CNAPs code.

xialeiliu commented 3 years ago

I was trying something else, at some point, it was conplaining about _simclr_episodefraction, so I added two more lines: EpisodeDescriptionConfig.ignore_hierarchy_probability = 0.0 EpisodeDescriptionConfig.simclr_episode_fraction = 0.0

but, when I rerun it again, it seems the code works perfectly fine.

It was a bit confusing for the meta_dataset_config.gin, it seems that most variables are not used, is it true?

jfb54 commented 3 years ago

Yes, most of the settings in meta_dataset_config.gin are over-ridden in the reader code. Agree, this is somewhat confusing.